A Review of Digital Innovations for Diet Monitoring and Precision Nutrition The central theme of this study is to propose more effective approaches for tracking diets, considering the limitations of current methods like food diaries and 24-hour recall. These conventional methods are burdensome and often result in inaccurate estimations of food consumption. Accurate diet monitoring holds significant importance as it has the potential to greatly reduce mortality rates and improve overall health through precise nutritional interventions. The first recommended method suggests leveraging mobile applications for diet monitoring. These apps serve as valuable complements to traditional food diaries, as users can conveniently log their dietary intake using their smartphones. These applications provide features that offer nutrient information for packaged foods and assist users in making informed decisions about meal choices and portion sizes. One notable advantage of these apps is the ability to track meals through photos, ensuring both convenience and accuracy. Users can capture real-time images of their meals, eliminating the need to rely on memory or enter multiple entries. Furthermore, the integration of photo diaries with artificial intelligence (AI) and continuous glucose monitoring (CGM) technologies allows for the identification of specific foods that may cause undesirable glucose spikes. The second proposed method involves the utilization of sensors to enhance the tracking of eating habits and nutritional intake, thereby increasing accuracy while reducing user burden. Two primary types of sensors are involved: physical and chemical sensors. Physical sensors detect specific gestures associated with eating, such as hand-to-mouth movements. While wearable sensors on wrists may be limited to laboratory settings, additional sensors like electromyography, piezoelectric, and acoustic sensors can be used to monitor muscle movements in the jaw, identifying chewing and swallowing sounds. Chemical sensors, on the other hand, employ dietary biomarkers to track nutritional content. Continuous Glucose Monitoring (CGM) is an example of such a biomarker, utilizing changes in blood glucose levels following a meal to provide insight into the macronutrient composition of consumed food. Through the application of CGM and machine learning, a study successfully developed a model capable of accurately predicting macronutrient components, showing promising results despite variations in individual food metabolism. Additionally, ketones derived from breath analysis serve as another biomarker, enabling the assessment of ketosis for individuals on ketogenic diets or individuals with diabetes at risk of ketoacidosis. Ongoing research is exploring the use of sweat and saliva to track nutritional components, although these approaches are still in the experimental stage. The final proposed method involves utilizing technology to develop personalized nutrition programs based on measurements of gut microbiome and blood glucose levels. A study conducted in this field focused on developing a machine learning model that utilizes CGM data to predict individual glucose responses to meals, incorporating factors such as microbiome composition and blood panel results. Companies can utilize the data generated by this model to formulate personalized nutrition programs, empowering users to effectively combat diabetes and other metabolic diseases. Overall, limitations include decreasing adherence over time, even with low-burden tools, and the potential impact on in-the-moment awareness, which can hinder weight loss. Another concern arises with personalized nutrition programs based on CGM, as measurements can vary across different devices. These limitations highlight the importance of engaging behavior-modification researchers to design interventions that promote adherence and lifestyle modification. Modeling Individual Differences in Food Metabolism through Altering Least Squares The motive of the study was to get a better understanding of the effects of macronutrients on blood glucose level, known as postprandial glucose response (PPGR). The blood glucose level is not only dependent on carbohydrates, but also dependent on protein, fat, and inter-individual differences in macronutrient metabolism. Understanding PPGR is crucial to real-world applications such as developing personalized nutrition programs and automatically monitoring diet using CGM. Due to the inter-individual differences in macronutrient metabolism, individuals have different PPGR after having the same meal. Previous works include using the shape of PPGR to predict the meal macronutrients’ amounts with a machine learning model, which was pretty successful given that the models were not customized for each participant. Another study built a machine learning model to predict PPGR based on “phenotypes” such as blood panels and gut microbiota. This research aims to learn how each macronutrient contributes to PPGR and capture the unique PPGR of each individual and use this information as a scaling factor for the impact of macronutrients. The researchers came up with an equation, X = A𝛼Z, where X is the PPGR, A is the basis function (the impact of macronutrients), 𝛼 is the sensitivity variable (varies by individual), and Z is macronutrients stored in each meal. They use a technique called altering least squares (ALS) to solve for A and 𝛼. They first assume 𝛼=1 to to estimate the matrix A, then solve for 𝛼 for each individual with value of A. Researchers discovered that increasing carbohydrates results in higher and prolonged PPGR peak while increasing protein and fat reduces the PPGR peak while making it more sustained. They also discovered that there is a wide range of sensitivity variables among the subject pool, meaning a large inter-individual difference in macronutrient metabolism, which makes the model based on averages rather limited. The model accounting for these sensitivity variables produced more effective results as for example, a subject who is more sensitive to fat has a weighted fat sensitivity variable for more accurate results of PPGR. The limitations of this research include an assumption that the macronutrients’ effects are linear and additive, which can be complemented in the future work where product terms and nonlinearities can be taken into account. Another direction for future research is using the sensitivity parameters to generate “synthetic patients” and develop personalized diet recommendations that reduce high glucose excursions after a meal. Towards the Development of Subject-independent Inverse Metabolic Models Diet monitoring is crucial in managing type 2 diabetes and other dietary diseases; however, traditional ways of monitoring diet are often time consuming and inaccurate. Using CGMs to automatically monitor dietary intake and using PPGR to estimate the macronutrient components of a meal could be more convenient and effective in managing the issue. Therefore, the researchers attempted to develop inverse metabolic models (IMM) to estimate the macronutrient components of a meal based on the shape of the PPGR. A lot of research has been done in terms of predicting the PPGR given the meal’s macronutrient components, but not vice versa. Previous research has discovered that there are huge inter-individual differences in the glucose response to a meal, which poses a great challenge when developing IMMs. They have developed machine learning models to predict the PPGR based on individual’s phenotypes such as blood panels and gut microbiota. The research used the signal processing methods which contain 4 steps: data preprocessing, feature extraction, standardization, and model training. First, they preprocessed raw PPGRs with a Kalman filter to denoise the signal and handle missing values. Next, they extracted features to capture PPGR shapes using Gaussian kernels. Next, they applied the standardization methods to reduce individual differences in PPGRs when training subject-independent models using three complementary approaches: baseline correction (accounts for the pre-meal glucose level), feature normalization (scale the range of a feature space relative to the minima/maxima of the data or their mean/standard deviation), and personalization (account for body composition of each person). Finally, they trained an IMM using gradient descent boosting (XGBoost) vs linear regression (LR) in order to predict the macronutrient composition of the meals from the resulting features. To evaluate the method, the researchers first compared XGBoost against LR without standardization; XGBoost outperformed LR, implying higher-dimensional and non-linear relationships between macronutrients and PPGRs. All three standardization methods complemented the inter-individual variability. For baseline correction, subtraction yielded better results than division. Feature-wise normalization improved carbohydrates and fat predictions over curve-wise normalization while protein predictions saw marginal improvements. For personalization, the researchers discovered that scaling the target values by BMI enhances both the correlation and RMSRE for all macronutrients. Limitations of the research include the controlled setting of the study, where participants were inactive after eating. This means that the model did not consider the impact of physical activity on lowering postprandial glucose levels. Therefore, future work is necessary to assess the effectiveness of our approach in more realistic environments and with a wider range of meal types. Currently, ongoing efforts involve conducting experiments where participants consume multiple solid and liquid meals throughout an extended period while engaging in their usual daily activities.